Wherefore art thou "researcher"?
The new role for researchers
Many claims about what AI can do are met with the a peanut gallery chorus of “yeah-buts.”
“Yeah, but it’s all hype.”
“Yeah, but it can’t get enough right.”
I’ve spent the last several months finding out for myself, and the moment it crystallized was mundane: I was at a coffee shop, on my phone, using Claude Code remotely to add purged walk-forward cross-validation to a backtesting pipeline, with an interface flexible enough to swap protocols quickly. Done before I finished the coffee.
The yeah-buts aren’t entirely wrong. They’re just aimed at the wrong part of the work.
The bigger build I’ve been iterating on is an agentic stock analysis workflow. Five pairs of analyst agents cover factor values, sector data, sentiment, technicals, and fundamentals, with each pair running competing bull and bear cases. An orchestrator coordinates them; a decision layer reconciles their outputs into buy, hold, or sell. Underneath it sits MARC, a research assistant with access to a 2,500-paper corpus embedded in a vector database, accessible through an MCP server. That infrastructure layer would have been a multi-month team project a few years ago. I built it in weeks.
The speed gain isn’t the interesting part. Where it comes from is.
The middle of the research workflow has compressed dramatically: writing code, wiring infrastructure, running sweeps. The two ends haven’t. Setting up the question, defining what a meaningful result looks like, deciding whether an apparent finding is real or an artifact — all of that takes the same time it always did. Sometimes more, because there’s now more output to sift.
So the researcher’s hours don’t disappear. They redistribute: less time in the middle, more at the start framing problems precisely enough for an agent to execute, more at the end scrutinizing what came back.
The end is where this matters most, because AI’s default disposition is generous. Here’s what that looks like in practice. Evaluating backtest results, a model would correctly label median Sharpe ratios as positive, but unless prompted, it wouldn’t exclude the ones that weren’t meaningfully different from zero. The result was a stream of presumptive positives, and when I looked under the hood, many were backed by median Sharpes indistinguishable from noise. A less suspicious reader gets a pipeline that manufactures “promising” strategies. Getting measured evaluation out of the model requires careful prompting and constant calibration. The judgment has to be supplied. The model executes against it.
That’s the pattern across the work. AI compresses tasks where the path from intent to output is mostly mechanical: syntax, scaffolding, well-specified transformations. It doesn’t compress knowing whether a sample is too small, whether a backtest is overfit, whether a result generalizes. Those calls take domain knowledge built over years, and they’re exactly the calls AI is least able to make on its own. If anything, that expertise is worth more now, because every error a researcher doesn’t catch compounds across far more output.
Fifteen years inside institutions also taught me what a solo setup can’t replicate, and why those things exist. Clean point-in-time data and reliable corporate actions are genuinely hard to assemble alone, and the gap shows up as caveats in any analysis. Risk review, peer challenge from colleagues who will push back hard on a thesis, capital to test ideas at scale: AI doesn’t remove those bottlenecks, and I wouldn’t want it to. What I’ve built is good for building, iterating, and learning.
Human-in-the-loop isn’t a token gesture. The human is more leveraged than before — one researcher can drive what used to take a small team — and the parts of the work that didn’t compress are the parts that determine whether the output is worth anything.
Until next time, stay on the cutting edge, everyone.